f4e2264a83
Copied from upstream_ref (NOT rewritten — exact upstream code):
ixformer C++ API (the authoritative header):
include/ixformer.h — ixformer::infer namespace: topk_softmax,
moe_compute_token_index_api, moe_w16a16_group_gemm, moe_expand_input,
moe_output_reduce_sum, silu_and_mul, rms_norm, xllm_paged_attention, etc.
include/ilu_ops_api.h — xllm::kernel::ilu namespace: moe_active_topk,
moe_gen_idx, moe_expand_input, group_gemm, moe_combine_result,
batch_prefill, batch_decode, rms_norm, matmul, act_and_mul, etc.
ILU kernel wrappers (call ixformer::infer directly):
csrc/ilu_kernel_fused_moe.cpp — topk routing + gen_idx + expand + combine
csrc/ilu_kernel_group_gemm.cpp — batched expert GEMM
csrc/ilu_kernel_{activation,norm,rope,matmul,attention}.cpp
ILU layer implementations (full pipeline):
csrc/ilu_layer_fused_moe.{cpp,h} — 797 lines, the complete MoE pipeline
that competitor 168 ran as corex_moe.py
csrc/ilu_layer_attention.{cpp,h} — prefill/decode attention dispatch
CUDA MoE kernels (from xllm + ds_vllm):
csrc/moe/moe_topk_softmax_kernels.cuh — CUB BlockReduce + warp topk
csrc/moe/moe_topk_sigmoid_kernels.cuh — sigmoid scoring variant
csrc/moe/moe_topk.cuh + moe_fused_topk.cu — entry points
csrc/moe/moeTopKFuncs.cuh — TRT-LLM derived vllm-compatible topk
csrc/moe/moe_ops.h + moe_align_sum_kernels.cu — alignment kernels
Common layer headers:
csrc/common_fused_moe{,_base}.h + common_moe_fused_topk.{cpp,h}
132 lines
4.4 KiB
C++
132 lines
4.4 KiB
C++
/* Copyright 2025 The xLLM Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
https://github.com/jd-opensource/xllm/blob/main/LICENSE
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#pragma once
|
|
|
|
#include <torch/torch.h>
|
|
|
|
#include "framework/model/model_args.h"
|
|
#include "framework/model/model_input_params.h"
|
|
#include "framework/parallel_state/parallel_args.h"
|
|
#include "framework/quant_args.h"
|
|
#include "framework/state_dict/state_dict.h"
|
|
#include "framework/state_dict/utils.h"
|
|
#include "layers/common/deep_ep.h"
|
|
#include "layers/common/dense_mlp.h"
|
|
#include "layers/common/fused_moe_base.h"
|
|
#include "layers/common/linear.h"
|
|
#include "platform/device.h"
|
|
#include "util/tensor_helper.h"
|
|
|
|
namespace xllm {
|
|
namespace layer {
|
|
|
|
class FusedMoEImpl : public torch::nn::Module {
|
|
public:
|
|
FusedMoEImpl() = default;
|
|
FusedMoEImpl(const ModelArgs& model_args,
|
|
const FusedMoEArgs& moe_args,
|
|
const QuantArgs& quant_args,
|
|
const ParallelArgs& parallel_args,
|
|
const torch::TensorOptions& options);
|
|
|
|
torch::Tensor forward_experts(const torch::Tensor& hidden_states,
|
|
const torch::Tensor& router_logits,
|
|
bool enable_all2all_communication);
|
|
torch::Tensor forward(const torch::Tensor& hidden_states,
|
|
const ModelInputParams& input_params);
|
|
void load_state_dict(const StateDict& state_dict);
|
|
|
|
private:
|
|
// struct to store the selected expert info
|
|
struct SelectedExpertInfo {
|
|
torch::Tensor reduce_weight;
|
|
torch::Tensor combine_idx;
|
|
torch::Tensor token_count_slice;
|
|
std::optional<torch::Tensor> cusum_token_count;
|
|
std::optional<torch::Tensor> input_scale;
|
|
};
|
|
|
|
// initial steps for MoE computation, select the experts for each token
|
|
torch::Tensor select_experts(const torch::Tensor& hidden_states_2d,
|
|
const torch::Tensor& router_logits_2d,
|
|
SelectedExpertInfo& selected_expert_info,
|
|
bool enable_all2all_communication);
|
|
|
|
private:
|
|
int64_t num_total_experts_;
|
|
int64_t topk_;
|
|
int64_t num_expert_group_;
|
|
int64_t topk_group_;
|
|
double route_scale_;
|
|
int64_t hidden_size_;
|
|
int64_t n_shared_experts_;
|
|
bool is_gated_;
|
|
int64_t renormalize_;
|
|
std::string hidden_act_;
|
|
std::string scoring_func_;
|
|
bool is_smoothquant_;
|
|
|
|
int64_t num_experts_per_rank_;
|
|
int64_t start_expert_id_;
|
|
|
|
// Deep EP related parameters
|
|
bool enable_deep_ep_;
|
|
DeepEPBuffer deep_ep_buffer_;
|
|
DeepEPParams deep_ep_params_;
|
|
torch::Tensor dispatch_recv_token_tensor_head_;
|
|
torch::Tensor dispatch_recv_token_tensor_tail_;
|
|
|
|
// steams for parallel shared experts
|
|
std::unique_ptr<Stream> shared_stream_;
|
|
std::unique_ptr<Stream> routed_stream_;
|
|
xllm::Device device_;
|
|
bool stream_initialized_ = false;
|
|
|
|
ReplicatedLinear gate_{nullptr};
|
|
DenseMLP shared_experts_{nullptr};
|
|
DeepEP deep_ep_{nullptr};
|
|
|
|
QuantArgs quant_args_;
|
|
ParallelArgs parallel_args_;
|
|
torch::TensorOptions options_;
|
|
ProcessGroup* tp_pg_;
|
|
|
|
DEFINE_WEIGHT(w13);
|
|
DEFINE_FUSED_WEIGHT(w1);
|
|
DEFINE_FUSED_WEIGHT(w3);
|
|
DEFINE_FUSED_WEIGHT(w2);
|
|
DEFINE_WEIGHT(e_score_correction_bias);
|
|
DEFINE_WEIGHT(w13_scale);
|
|
DEFINE_FUSED_WEIGHT(w1_scale);
|
|
DEFINE_FUSED_WEIGHT(w3_scale);
|
|
DEFINE_FUSED_WEIGHT(w2_scale);
|
|
DEFINE_FUSED_WEIGHT(input_smooth);
|
|
DEFINE_FUSED_WEIGHT(act_smooth);
|
|
|
|
void load_e_score_correction_bias(const StateDict& state_dict);
|
|
void load_experts(const StateDict& state_dict);
|
|
// create the group gemm output tensor with the workspace
|
|
torch::Tensor create_group_gemm_output(const torch::Tensor& a,
|
|
const torch::Tensor& b,
|
|
const torch::Tensor& group_list,
|
|
torch::ScalarType dtype,
|
|
torch::Tensor& workspace);
|
|
};
|
|
TORCH_MODULE(FusedMoE);
|
|
|
|
} // namespace layer
|
|
} // namespace xllm
|